🚨 🚨 Inherited CausalLM Tests (#37590)
* stash commit * Experiment 1: Try just Gemma * Experiment 1: Just try Gemma * make fixup * Trigger tests * stash commit * Try adding Gemma3 as well * make fixup * Correct attrib names * Correct pipeline model mapping * Add in all_model_classes for Gemma1 again * Move the pipeline model mapping around again * make fixup * Revert Gemma3 changes since it's a VLM * Let's try Falcon * Correct attributes * Correct attributes * Let's try just overriding get_config() for now * Do Nemotron too * And Llama! * Do llama/persimmon * Correctly skip tests * Fix Persimmon * Include Phimoe * Fix Gemma2 * Set model_tester_class correctly * Add GLM * More models! * models models models * make fixup * Add Qwen3 + Qwen3MoE * Correct import * make fixup * Add the QuestionAnswering classes * Add the QuestionAnswering classes * Move pipeline mapping to the right place * Jetmoe too * Stop RoPE testing models with no RoPE * Fix up JetMOE a bit * Fix up JetMOE a bit * Can we just force pad_token_id all the time? * make fixup * fix starcoder2 * Move pipeline mapping * Fix RoPE skipping * Fix RecurrentGemma tests * Fix Falcon tests * Add MoE attributes * Fix values for RoPE testing * Make sure we set bos_token_id and eos_token_id in an appropriate range * make fixup * Fix GLM4 * Add mamba attributes * Revert bits of JetMOE * Re-add the JetMOE skips * Update tests/causal_lm_tester.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Add licence --------- Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
This commit is contained in:
Matt
@@ -16,20 +16,14 @@
|
||||
|
||||
import unittest
|
||||
|
||||
from parameterized import parameterized
|
||||
|
||||
from transformers import PhimoeConfig, StaticCache, is_torch_available, set_seed
|
||||
from transformers import PhimoeConfig, StaticCache, is_torch_available
|
||||
from transformers.testing_utils import (
|
||||
is_flaky,
|
||||
require_torch,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
|
||||
from ...generation.test_utils import GenerationTesterMixin
|
||||
from ...test_configuration_common import ConfigTester
|
||||
from ...test_modeling_common import ModelTesterMixin, ids_tensor
|
||||
from ...test_pipeline_mixin import PipelineTesterMixin
|
||||
from ...causal_lm_tester import CausalLMModelTest, CausalLMModelTester
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
@@ -92,138 +86,23 @@ if is_torch_available():
|
||||
return response_tokens
|
||||
|
||||
|
||||
class PhimoeModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_input_mask=True,
|
||||
use_token_type_ids=False,
|
||||
use_labels=True,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=2,
|
||||
num_attention_heads=4,
|
||||
num_key_value_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=131072,
|
||||
type_vocab_size=16,
|
||||
type_sequence_label_size=2,
|
||||
initializer_range=0.02,
|
||||
num_labels=3,
|
||||
num_choices=4,
|
||||
pad_token_id=0,
|
||||
scope=None,
|
||||
original_max_position_embeddings=4096,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.use_input_mask = use_input_mask
|
||||
self.use_token_type_ids = use_token_type_ids
|
||||
self.use_labels = use_labels
|
||||
self.vocab_size = vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.num_key_value_heads = num_key_value_heads
|
||||
self.intermediate_size = intermediate_size
|
||||
self.hidden_act = hidden_act
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.type_vocab_size = type_vocab_size
|
||||
self.type_sequence_label_size = type_sequence_label_size
|
||||
self.initializer_range = initializer_range
|
||||
self.num_labels = num_labels
|
||||
self.num_choices = num_choices
|
||||
self.pad_token_id = pad_token_id
|
||||
self.scope = scope
|
||||
self.original_max_position_embeddings = original_max_position_embeddings
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs
|
||||
def prepare_config_and_inputs(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
input_mask = None
|
||||
if self.use_input_mask:
|
||||
input_mask = torch.tril(torch.ones_like(input_ids).to(torch_device))
|
||||
|
||||
token_type_ids = None
|
||||
if self.use_token_type_ids:
|
||||
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
|
||||
|
||||
sequence_labels = None
|
||||
token_labels = None
|
||||
choice_labels = None
|
||||
if self.use_labels:
|
||||
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
|
||||
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
||||
choice_labels = ids_tensor([self.batch_size], self.num_choices)
|
||||
|
||||
config = self.get_config()
|
||||
|
||||
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
||||
|
||||
def get_config(self):
|
||||
return PhimoeConfig(
|
||||
vocab_size=self.vocab_size,
|
||||
hidden_size=self.hidden_size,
|
||||
num_hidden_layers=self.num_hidden_layers,
|
||||
num_attention_heads=self.num_attention_heads,
|
||||
num_key_value_heads=self.num_key_value_heads,
|
||||
intermediate_size=self.intermediate_size,
|
||||
hidden_act=self.hidden_act,
|
||||
hidden_dropout_prob=self.hidden_dropout_prob,
|
||||
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
type_vocab_size=self.type_vocab_size,
|
||||
is_decoder=False,
|
||||
initializer_range=self.initializer_range,
|
||||
pad_token_id=self.pad_token_id,
|
||||
num_experts_per_tok=2,
|
||||
num_local_experts=2,
|
||||
original_max_position_embeddings=self.original_max_position_embeddings,
|
||||
)
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model with Llama->Phimoe
|
||||
def create_and_check_model(
|
||||
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
||||
):
|
||||
model = PhimoeModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=input_mask)
|
||||
result = model(input_ids)
|
||||
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config_and_inputs = self.prepare_config_and_inputs()
|
||||
(
|
||||
config,
|
||||
input_ids,
|
||||
token_type_ids,
|
||||
input_mask,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
) = config_and_inputs
|
||||
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
|
||||
return config, inputs_dict
|
||||
class PhimoeModelTester(CausalLMModelTester):
|
||||
if is_torch_available():
|
||||
config_class = PhimoeConfig
|
||||
base_model_class = PhimoeModel
|
||||
causal_lm_class = PhimoeForCausalLM
|
||||
sequence_class = PhimoeForSequenceClassification
|
||||
|
||||
|
||||
@require_torch
|
||||
class PhimoeModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
||||
class PhimoeModelTest(CausalLMModelTest, unittest.TestCase):
|
||||
all_model_classes = (
|
||||
(PhimoeModel, PhimoeForCausalLM, PhimoeForSequenceClassification) if is_torch_available() else ()
|
||||
)
|
||||
|
||||
test_headmasking = False
|
||||
test_pruning = False
|
||||
model_tester_class = PhimoeModelTester
|
||||
pipeline_model_mapping = (
|
||||
{
|
||||
"feature-extraction": PhimoeModel,
|
||||
@@ -235,150 +114,12 @@ class PhimoeModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMix
|
||||
else {}
|
||||
)
|
||||
|
||||
test_headmasking = False
|
||||
test_pruning = False
|
||||
|
||||
# TODO (ydshieh): Check this. See https://app.circleci.com/pipelines/github/huggingface/transformers/79292/workflows/fa2ba644-8953-44a6-8f67-ccd69ca6a476/jobs/1012905
|
||||
def is_pipeline_test_to_skip(
|
||||
self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
|
||||
):
|
||||
return True
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.setUp with Llama->Phimoe
|
||||
def setUp(self):
|
||||
self.model_tester = PhimoeModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=PhimoeConfig, hidden_size=37)
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_config
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_model
|
||||
def test_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_model(*config_and_inputs)
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_sequence_classification_model with Llama->Phimoe,llama->phimoe
|
||||
def test_phimoe_sequence_classification_model(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
config.num_labels = 3
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
|
||||
model = PhimoeForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
|
||||
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_sequence_classification_model_for_single_label with Llama->Phimoe,llama->phimoe
|
||||
def test_phimoe_sequence_classification_model_for_single_label(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
config.num_labels = 3
|
||||
config.problem_type = "single_label_classification"
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
|
||||
model = PhimoeForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
|
||||
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_sequence_classification_model_for_multi_label with Llama->Phimoe,llama->phimoe
|
||||
def test_phimoe_sequence_classification_model_for_multi_label(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
config.num_labels = 3
|
||||
config.problem_type = "multi_label_classification"
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
sequence_labels = ids_tensor(
|
||||
[self.model_tester.batch_size, config.num_labels], self.model_tester.type_sequence_label_size
|
||||
).to(torch.float)
|
||||
model = PhimoeForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
|
||||
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
|
||||
|
||||
@parameterized.expand([("longrope",)])
|
||||
def test_model_rope_scaling_from_config(self, scaling_type):
|
||||
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
short_input = ids_tensor([1, 10], config.vocab_size)
|
||||
long_input = ids_tensor([1, int(config.original_max_position_embeddings * 1.5)], config.vocab_size)
|
||||
|
||||
set_seed(42) # Fixed seed at init time so the two models get the same random weights
|
||||
original_model = PhimoeModel(config)
|
||||
original_model.to(torch_device)
|
||||
original_model.eval()
|
||||
original_short_output = original_model(short_input).last_hidden_state
|
||||
original_long_output = original_model(long_input).last_hidden_state
|
||||
|
||||
set_seed(42) # Fixed seed at init time so the two models get the same random weights
|
||||
n_factors = config.hidden_size // config.num_attention_heads // 2
|
||||
config.rope_scaling = {
|
||||
"type": scaling_type,
|
||||
"short_factor": [3.0 for _ in range(n_factors)],
|
||||
"long_factor": [5.0 for _ in range(n_factors)],
|
||||
"short_mscale": 1.243163121016122,
|
||||
"long_mscale": 1.243163121016122,
|
||||
"original_max_position_embeddings": 4096,
|
||||
}
|
||||
scaled_model = PhimoeModel(config)
|
||||
scaled_model.to(torch_device)
|
||||
scaled_model.eval()
|
||||
scaled_short_output = scaled_model(short_input).last_hidden_state
|
||||
scaled_long_output = scaled_model(long_input).last_hidden_state
|
||||
|
||||
# Scaling changes the RoPE embeddings, both for the short and long outputs
|
||||
self.assertFalse(torch.allclose(original_short_output, scaled_short_output, atol=1e-5))
|
||||
self.assertFalse(torch.allclose(original_long_output, scaled_long_output, atol=1e-5))
|
||||
|
||||
@parameterized.expand([("longrope",)])
|
||||
@is_flaky() # TODO (joao): unify rope tests in the mixin
|
||||
def test_model_rope_scaling_short_long_factor(self, scaling_type):
|
||||
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
n_factors = config.hidden_size // config.num_key_value_heads // 2
|
||||
config.rope_scaling = {
|
||||
"type": scaling_type,
|
||||
"short_factor": [3.0 for _ in range(n_factors)],
|
||||
"long_factor": [5.0 for _ in range(n_factors)],
|
||||
"short_mscale": 1.243163121016122,
|
||||
"long_mscale": 1.243163121016122,
|
||||
"original_max_position_embeddings": 4096,
|
||||
}
|
||||
input_tensor = ids_tensor([1, 4090], config.vocab_size)
|
||||
model = PhimoeForCausalLM(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
generation_args_short = {
|
||||
"max_length": config.original_max_position_embeddings,
|
||||
"temperature": 0.0,
|
||||
"use_cache": True,
|
||||
"do_sample": False,
|
||||
"return_dict_in_generate": True,
|
||||
}
|
||||
output_with_short_factor = model.generate(input_tensor, **generation_args_short)
|
||||
keys_with_short_factor = output_with_short_factor.past_key_values[0][0]
|
||||
generation_args_long = {
|
||||
"max_length": config.original_max_position_embeddings + 5,
|
||||
"temperature": 0.0,
|
||||
"use_cache": True,
|
||||
"do_sample": False,
|
||||
"return_dict_in_generate": True,
|
||||
"output_logits": True,
|
||||
}
|
||||
output_with_long_factor = model.generate(input_tensor, **generation_args_long)
|
||||
keys_with_long_factor = output_with_long_factor.past_key_values[0][0]
|
||||
last_token_logits = output_with_long_factor.logits[-1][-1]
|
||||
regenerated_last_token_logits = model(output_with_long_factor.sequences[:, :-1]).logits[0][-1]
|
||||
keys_with_long_factor = keys_with_long_factor[:, :, : config.original_max_position_embeddings - 1, :]
|
||||
|
||||
# KV cache is re-computed after reaching the (`config.original_max_position_embeddings`+1)th token position
|
||||
self.assertFalse(torch.allclose(keys_with_short_factor, keys_with_long_factor, atol=1e-3, rtol=1e-3))
|
||||
# Last token generated using long factor
|
||||
torch.testing.assert_close(last_token_logits, regenerated_last_token_logits, rtol=1e-2, atol=1e-2)
|
||||
|
||||
|
||||
@slow
|
||||
@require_torch
|
||||
|
||||
Reference in New Issue
Block a user